| The traditional fixed road monitoring method has some problems such as the timeliness of feedback information and the blind area of monitoring scope,which can not meet the increasingly developing requirements of road traffic management.UAV is flexible,light and portable,and can be conveniently equipped with camera for aerial vehicle detection,which has become an important development direction and research hotspot in the new intelligent traffic management system.Due to the complexity of the actual road traffic environment changes,the drones influence on image quality,and small object detection in alcohol content is not high and the close distance between vehicles and residual position location is not accurate,this paper has carried out based on the depth study of uav aerial image detection and recognition research,the main research content is as follows:For hilly domain traffic environment there are mass of fog and low illumination,affect the quality of aerial images and put forward a kind of to fog algorithm based on AOD-Net,using convolution neural network is used to estimate values of K,avoid error amplifier through computing intermediate parameters affect the image to fog effect,improves the speed of the fogging algorithm to improve the image to fog fog effect;For low illumination images,a method based on the combination of improved histogram equalization and Laplacian sharpening is adopted.Finally,experiments verify that the image preprocessing algorithm designed in this paper can improve the accuracy of object detection algorithm to a certain extent without affecting the detection speed of aerial photography.For unmanned aerial vehicle(UAV)for traffic monitoring aerial,the traditional methods for detecting small targets in the images of the vehicle has the problem of accuracy is not high,in Darknet YOLOv3 network backbone structure-53 append a final residual module 1 * 1 and 3 * 3 convolution layer,on the premise of no loss of resolution increase nonlinear excitation,to increase the information interaction between the channel capacity,improve the ability of expression;In order to improve the detection accuracy of small targets,the 36 th layer and the 11 th layer of the network are spliced and fused on the premise of retaining the sampling detection of the original three layers.YOLOv3 algorithms of image in the close distance between vehicles,which can identify easily mistaken for the same car to leak,this paper proposes a prediction based on the maximum inhibition box generation algorithm,it will all candidate box with confidence after box candidates were occurring simultaneously with the highest scores than will be greater than a certain threshold candidate box no zero directly,but through gaussian weighted,confidence that box to the candidate to structure optimization.Aiming at the problem of vehicle positioning accuracy is not high,the introduction of normal distribution in the object detection of network output,the standard deviation produced by normal distribution as positioning confidence to improve the accuracy of the test frame orientation,at the same time adopted a candidate based on anchor box generation algorithm,through the method of maximum minimum distance product clustering center to initialize clustering is sensitive to the initial value of the problem to reduce,thus to improve the detection precision of the bounding box,in order to more accurate lock vehicle location.Finally,the improved network was built and trained based on the Darknet source framework,and the weight file after training was obtained.Then,the improved network was used to conduct experiments on the images collected by UAV aerial photography.The results show that,compared with the previous detection algorithms,the object detection algorithm based on complex environment proposed in this paper improves the whole detection accuracy by 2.25% on the premise of meeting the real-time requirements of road monitoring.It has good performance of detection and recognition. |
PDF Full Text Request